XWind router slashes AI inference latency by 52% using wind energy

The paper introduces AI Greenferencing, a deployment model that colocates modular AI compute clusters directly at renewable energy farms—starting with wind. This behind-the-meter approach generates local demand for renewable energy, reduces transmission losses, and sidesteps the high cost and long timelines of grid expansion. The authors estimate that over 890 GW of wind capacity is within 50ms network round-trip time of Azure data centers, making it feasible to serve inference traffic without violating latency budgets.

To handle the variability of wind power, the team built XWind, a lightweight, reactive inference router. It uses only real-time signals—inference latency, KV-cache utilization, and queue depth—to dynamically activate or deactivate sites and route requests. Tested on a real 64-GPU A100 cluster emulating three wind-powered sites with production Azure LLM traces, XWind achieved up to 52% lower P99 end-to-end latency than the strongest competing approach (also the authors' prior work) and up to 98% lower than baselines such as power-capping and GPU idling. Gains were consistent across workloads, load levels, and GPU generations, suggesting a practical path to greener, more resilient AI infrastructure.